Skip to content

Fix a miscompile that occurs when reading a matrix contained in a struct with no sibling fields #8568

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
aclysma wants to merge 2 commits into
base: main
Choose a base branch
from
+55 −11
Open

Fix a miscompile that occurs when reading a matrix contained in a struct with no sibling fields #8568

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
748e030
Fix a miscompile that occurs when reading a matrix contained in a str…
aclysma Jun 19, 2026
d6a4e6b
clang format fix
aclysma Jun 19, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
35 changes: 24 additions & 11 deletions lib/HLSL/HLOperationLower.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -9222,14 +9222,21 @@ void TranslateStructBufMatSubscript(CallInst *CI, Value *handle,
assert(resultSize <= 16);
std::vector<Value *> idxList(resultSize);

// For raw buffers the byte offset rides in the buffer index and the element
// offset operand stays undef, so the per-element offset must accumulate into
// bufIdx. For structured buffers the per-element offset is the element offset
// and bufIdx is the structure index.
bool isRawBuf = DXIL::IsRawBuffer(ResKind);
Value *matBaseIdx = isRawBuf ? bufIdx : baseOffset;

switch (subOp) {
case HLSubscriptOpcode::ColMatSubscript:
case HLSubscriptOpcode::RowMatSubscript: {
for (unsigned i = 0; i < resultSize; i++) {
Value *offset =
CI->getArgOperand(HLOperandIndex::kMatSubscriptSubOpIdx + i);
offset = subBuilder.CreateMul(offset, EltByteSize);
idxList[i] = subBuilder.CreateAdd(baseOffset, offset);
idxList[i] = subBuilder.CreateAdd(matBaseIdx, offset);
}
} break;
case HLSubscriptOpcode::RowMatElement:
Expand All @@ -9238,7 +9245,7 @@ void TranslateStructBufMatSubscript(CallInst *CI, Value *handle,
for (unsigned i = 0; i < resultSize; i++) {
Value *offset =
subBuilder.CreateMul(EltIdxs->getAggregateElement(i), EltByteSize);
idxList[i] = subBuilder.CreateAdd(baseOffset, offset);
idxList[i] = subBuilder.CreateAdd(matBaseIdx, offset);
}
} break;
default:
Expand All @@ -9252,17 +9259,19 @@ void TranslateStructBufMatSubscript(CallInst *CI, Value *handle,
Value *subsUser = *(U++);
if (resultSize == 1) {
TranslateStructBufSubscriptUser(cast<Instruction>(subsUser), handle,
ResKind, bufIdx, idxList[0], status,
hlslOP, DL);
ResKind, isRawBuf ? idxList[0] : bufIdx,
isRawBuf ? baseOffset : idxList[0],
status, hlslOP, DL);
continue;
}
if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(subsUser)) {
Value *GEPOffset = LowerGEPOnMatIndexListToIndex(GEP, idxList);

for (auto gepU = GEP->user_begin(); gepU != GEP->user_end();) {
Instruction *gepUserInst = cast<Instruction>(*(gepU++));
TranslateStructBufSubscriptUser(gepUserInst, handle, ResKind, bufIdx,
GEPOffset, status, hlslOP, DL);
TranslateStructBufSubscriptUser(
gepUserInst, handle, ResKind, isRawBuf ? GEPOffset : bufIdx,
isRawBuf ? baseOffset : GEPOffset, status, hlslOP, DL);
}

GEP->eraseFromParent();
Expand All @@ -9276,13 +9285,15 @@ void TranslateStructBufMatSubscript(CallInst *CI, Value *handle,
for (unsigned i = 0; i < resultSize; i++) {
Value *EltVal = stBuilder.CreateExtractElement(Val, i);
uint8_t mask = DXIL::kCompMask_X;
GenerateStructBufSt(handle, bufIdx, idxList[i], EltTy, hlslOP,
GenerateStructBufSt(handle, isRawBuf ? idxList[i] : bufIdx,
isRawBuf ? baseOffset : idxList[i], EltTy, hlslOP,
stBuilder, {EltVal, undefElt, undefElt, undefElt},
mask, alignment);
}
} else {
uint8_t mask = DXIL::kCompMask_X;
GenerateStructBufSt(handle, bufIdx, idxList[0], EltTy, hlslOP,
GenerateStructBufSt(handle, isRawBuf ? idxList[0] : bufIdx,
isRawBuf ? baseOffset : idxList[0], EltTy, hlslOP,
stBuilder, {Val, undefElt, undefElt, undefElt},
mask, alignment);
}
Expand All @@ -9300,14 +9311,16 @@ void TranslateStructBufMatSubscript(CallInst *CI, Value *handle,
for (unsigned i = 0; i < resultSize; i++) {
Value *ResultElt;
// TODO: This can be inefficient for row major matrix load
GenerateRawBufLd(handle, bufIdx, idxList[i],
GenerateRawBufLd(handle, isRawBuf ? idxList[i] : bufIdx,
isRawBuf ? baseOffset : idxList[i],
/*status*/ nullptr, EltTy, ResultElt, hlslOP,
ldBuilder, 1, alignment);
ldData = ldBuilder.CreateInsertElement(ldData, ResultElt, i);
}
} else {
GenerateRawBufLd(handle, bufIdx, idxList[0], /*status*/ nullptr, EltTy,
ldData, hlslOP, ldBuilder, 4, alignment);
GenerateRawBufLd(handle, isRawBuf ? idxList[0] : bufIdx,
isRawBuf ? baseOffset : idxList[0], /*status*/ nullptr,
EltTy, ldData, hlslOP, ldBuilder, 4, alignment);
}
ldUser->replaceAllUsesWith(ldData);
ldUser->eraseFromParent();
Expand Down
31 changes: 31 additions & 0 deletions .../HLSLFileCheck/hlsl/objects/ByteAddressBuffer/bab_load_struct_matrix_element_offsets.hlsl
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,31 @@
// Reading individual matrix elements through the struct member of a
// ByteAddressBuffer.Load<struct{float4x4}> must address each element at its own
// byte offset. For raw buffers the byte offset rides in the load index operand
// and the element-offset operand stays undef, so the four first-column reads
// below must land at indices 0, 4, 8 and 12 -- they must not collapse to the
// matrix base (index 0), which is what happened before the matrix-element
// subscript path was taught the raw-buffer addressing convention.

// RUN: %dxc -T cs_6_6 -E cs %s | FileCheck %s

struct Box { float4x4 m; };

ByteAddressBuffer src : register(t0);
RWByteAddressBuffer dst : register(u0);

[numthreads(1, 1, 1)]
void cs()
{
Box b = src.Load<Box>(0);

// Column-major: _m00/_m10/_m20/_m30 live at bytes 0/4/8/12.
dst.Store(0, asuint(b.m._m00));
dst.Store(4, asuint(b.m._m10));
dst.Store(8, asuint(b.m._m20));
dst.Store(12, asuint(b.m._m30));
}

// CHECK: rawBufferLoad.f32(i32 139, %dx.types.Handle %{{.*}}, i32 0, i32 undef, i8 1, i32 4)
// CHECK: rawBufferLoad.f32(i32 139, %dx.types.Handle %{{.*}}, i32 4, i32 undef, i8 1, i32 4)
// CHECK: rawBufferLoad.f32(i32 139, %dx.types.Handle %{{.*}}, i32 8, i32 undef, i8 1, i32 4)
// CHECK: rawBufferLoad.f32(i32 139, %dx.types.Handle %{{.*}}, i32 12, i32 undef, i8 1, i32 4)
Loading